Few-shot fine-grained image classification involves using limited samples to classify images from novel subcategories within the same category. Recent research indicates that classifiers based on reconstruction for few-shot learning attain elevated accuracy levels due to their capacity to preserve greater detail in appearance. However, they reconstruct using a weighted sum of all local descriptors and consider the reconstruction error of all descriptors for classification. This may lead to the reconstruction and utilization of task-irrelevant descriptors for classification, potentially misguiding the final outcome. This study presents, for the initial instance, a task-aware discriminative local descriptors reconstruction mechanism to address these issues, which can adaptively filter out task-irrelevant descriptors throughout the task, selecting highly discriminative descriptors for reconstruction. This design effectively aids the model in filtering out redundant information and exploring more nuanced and distinctive features throughout the task. Additionally, our unique detail-aware self-reconstruction module further refines feature discriminability. Extensive experimental results on fine-grained and generalized datasets consistently demonstrate that the proposed TARNet surpasses current state-of-the-art methods.